Research prototype: multimodal RAG for financial filings. Hybrid dense+BM25 retrieval (RRF), numeric grounding checks, a sandboxed calculator and cited answers. Measured on a 119-question golden set: evidence hit@5 0.71, asked-for figure stated in 68% of answers (one model). Chart extraction is implemented but not yet validated. FastAPI, CLI.
-
Updated
Nov 21, 2026 - Python